Aspects of Automation of Selective Cleaning
Faculty of Forest Sciences Department of Silviculture
Swedish University of Agricultural Sciences
Acta Universitatis Agriculturae Sueciae2005:74
Vestlund, K. 2005. Aspects of automation of selective cleaning. Doctor’s dissertation.
ISSN 1652-6880, ISBN 91-576-6973-2
Cleaning (pre-commercial thinning) is a silvicultural operation, primarily used to improve growing conditions of remaining trees in young stands (ca. 3 - 5 m of height). Cleaning costs are considered high in Sweden and the work is laborious. Selective cleaning with autonomous artificial agents (robots) may rationalise the work, but requires new knowledge. This thesis aims to analyse key issues regarding automation of cleaning;
suggesting general solutions and focusing on automatic selection of main-stems. The essential requests put on cleaning robots are to render acceptable results and to be cost competitive. They must be safe and be able to operate independently and unattended for several hours in a dynamic and non-deterministic environment. Machine vision, radar, and laser scanners are promising techniques for obstacle avoidance, tree identification, and tool control. Horizontal laser scannings were made, demonstrating the possibility to find stems and make estimations regarding their height and diameter. Knowledge regarding stem selections was retrieved through qualitative interviews with persons performing cleaning.
They consider similar attributes of trees, and these findings and current cleaning manuals were used in combination with a field inventory in the development of a decision support system (DSS). The DSS selects stems by the attributes species, position, diameter, and damage. It was used to run computer-based simulations in a variety of young forests. A general follow-up showed that the DSS produced acceptable results. The DSS was further evaluated by comparing its selections with those made by experienced cleaners, and by a test in which laymen performed cleanings following the system. The DSS seems to be useful and flexible, since it can be adjusted in accordance with the cleaners’ results. The laymen’s results implied that the DSS is robust and that it could be used as a training tool.
Using the DSS in automatic, or semi-automatic, cleaning operations should be possible if and when selected attributes can be automatically perceived. A suitable base-machine and thorough research, regarding e.g. safety, obstacle avoidance, and target identification, is needed to develop competitive robots. However, using the DSS as a training-tool for inexperienced cleaners could be an interesting option as of today.
Key words: autonomous off-road vehicle, decision support system, interviews, robot, pre- commercial thinning, simulations, training-tool.
Author’s address: Karin Vestlund, Department of Silviculture, Swedish University of Agricultural Sciences, P.O. Box 7060, SE-750 07 UPPSALA, Sweden.
Introduction, 7 Cleaning, 7
Development of forestry operations in Sweden, 8 Automatic selective cleaning, 11
Material and methods, 15 Review, 15
The Decision Support System, 15 Field inventory, 16
Field “experiments” with cleaners, 19 Field “experiments” with laymen, 20 Laser scanning, 20
Statistics, 20 Results, 22 Papers I and V, 22 Paper II, 22 Paper III, 23 Paper IV, 25 Discussion, 26
Research approach, 26
Prospect of a cleaning robot, 29 Changing the system, 31
Prospect of automatic selection of stems, 33 Alternative approaches for future cleanings, 37 Future research, 39
Conclusions, 41 References, 42
Acknowledgements, 51 Swedish summary, 52
This thesis is based on the following papers, which will be referred to by their Roman numerals:
I. Vestlund, K. & Hellström, T. Requirements and system design for a robot performing selective cleaning in young forest stands. (Accepted for publication in Journal of Terramechanics).
II. Vestlund, K. 2004. Assessing rules and ideas for stem selection in cleaning.
Baltic Forestry, 10(2): 61-71.
III. Vestlund, K., Nordfjell, T., Eliasson, L. & Karlsson, A. A decision support system for selective cleaning. (Submitted).
IV. Vestlund, K., Nordfjell, T. & Eliasson, L. Comparison of human and computer-based selective cleaning. (Submitted).
V. Erikson, M. & Vestlund, K. 2003. Finding tree-stems in laser range images of young mixed stands to perform selective cleaning. In: Hyyppä, J., Naesset, E., Olsson, H., Granqvist-Pahlén, T. & Reese, H. (Eds.) Proceedings of the Scandlaser scientific workshop on airborne laser scanning of forests. Working paper 112, Department of Forest Resource, Management and Geomatics, Swedish University of Agricultural Sciences, Umeå, Sweden. p. 244-250. ISSN: 1401-1204.
Paper I, II, and V are printed/reproduced with the kind permission of the publishers.
The forest is valuable for Sweden as a whole. In 2003 the export value of forest and forest industry products was 109 billion SEK, which was 13% of Sweden’s total export value (Anon. 2004a). Thus, it is important to have progress in all aspects of forestry.
Since 1905, it has been compulsory in Sweden to reforest harvested areas (Holmberg 2005). The common silvicultural practice from the late 1950s has been a sequence of clear-cutting followed by site-preparation, planting or natural regeneration, and then tending of the stand during the whole rotation period until a new harvest can be performed, i.e. even-age management. Standard tending operations in Swedish forests include cleaning (pre-commercial thinning) and thinning (commercial thinning) (cf. Anon. 2004a).
Cleaning is performed in young forest stands, prior to thinnings, in Nordic countries usually when the stand is ca. three to five meters of height (Varmola &
Salminen 2004). The National Board of Forestry defines cleaning as the thinning of a stand, in which the main part of the cut volume originates from stems of less than 10 cm in diameter at breast height (Pettersson & Bäcke 1998). In dense stands a reduction in the number of stems increases the volume growth per stem (Aussenac & Granier 1988, Pettersson 1993), as the remaining trees (main-stems) can benefit from increased access to nutrients, water, and light (Eriksson 1976).
The most common reason (the broad objective) for performing cleanings mentioned in Sweden is economic; concerning increased revenue later in the rotation period (Berg et al. 1973) and/or reduced future silvicultural costs (Håkansson & Steffen 1994). To achieve the broad-objective usual sub-objectives of cleaning are to produce more volume per stem, to produce stems with certain characteristics, and to minimise the operational costs. Thus compromises between sub-objectives might be needed if/when improvement for one sub-objective only can be accomplished at the expense of another (cf. Keeney & Raiffa 1993).
The lower the height and the lower the number of stems per hectare the faster cleanings can be performed (Bergstrand et al. 1986), so cleaning young stands should be cheaper than cleaning older stands. However, premature cleanings can have disadvantageous effects, such as rank-growth of trees (causing e.g. thick branches) or establishment of sprouts causing a new demand for cleaning (Eriksson 1976, Andersson 1985). Postponed cleanings can result in over-dense stands, with high risks for damage by snow, insects, wind etc. (Eriksson 1976).
Furthermore, no or very low-intensity* cleanings increase the share of early
thinnings with a negative revenue and the risk of getting a lower income over the whole rotation-period because the growth is divided on a large number of stems (Frohm 1996). Cleaning manuals state that the operation should be performed when there is a risk that potential main-stems will be adversely affected by competition and/or damaged (cf. Karlsson et al. 1997). Cleaning manuals for coniferous (Scots pine and Norway spruce) stands in Sweden of which the average height is some three meters state that, depending on site quality and species, between 1 400 and 4 000 stems per hectare should remain (e.g. Pettersson &
Bäcke 1998, Anon. 1999a, Normark & Bergqvist 2000). In Finland and Norway the recommended spacing after cleaning of Scots pine stands varies from 2 000 to 3 500 stems per hectare (Varmola & Salminen 2004).
Cleaning operations can be selective, geometrical, or a combination of both (Berg et al. 1973). Geometrical cleaning can be cheaper than selective cleaning in stands with at least 10-20 000 stems per hectare (Fryk 1985, Ryans 1988, Bergkvist &
Nordén 2004). Reasons for making individual selections include a desire to improve stand quality and/or influence species composition, which might increase the final profitability (e.g. Berg et al. 1973). Selective cleaning is frequently used today in many parts of the world (cf. Anon. 1999b, Kaivola 1996, Strobl & Bell 2000, Ek 2003, Ladrach 2004). Geometrical cleaning, in strip, checkerboard, or other patterns, is common in loblolly pine (Pinus taeda L.) stands in USA (Lloyd
& Waldrop 1999) and also used in Canada (Ryans 1988, Ryans & St-Amour 1996) and in dense natural generations of beech in Denmark (Möller-Madsen &
Petersen 2002). Chemical cleaning can be an inexpensive approach for reducing the amount of deciduous stems, but it is not widely used due to its environmental affects. However, herbicides are used in some 35% of the treated area in Canada (Ryans & St-Amour 1996, Anon. 2004b).
Consequently, it is generally difficult to make decisions such as when and how to clean, see also the section Automatic selective cleaning.
Development of forestry operations in Sweden
Forestry has undergone considerable changes in the Nordic countries during the last century, from a labour intensive to a capital-intensive business. In the late 19th century, the industrial revolution reached Swedish forests and along the riverbanks and the coastline sawmills and later on pulp mills were established. This was by reason of the demand for sawn timber in Europe. The forest operations at that time were highly seasonal; the round wood was harvested during the winter with axes and saws, and transported to rivers or lakes by horse-drawn sleighs (Kardell 2004). The transportation of logs to the mills took place in the spring as the rivers were used for floating the logs downstream. Cleaning was introduced into Swedish forestry during this century using manual tools like brash hooks, special shears, axes, saws and knifes (Björkman 1877, Wahlgren 1914).
and in company owned forests by 1936 (Carpelan 1948). In the middle of the 1950s most harvesting was performed with power-saws (Andersson 1986) and motor-manual brush saws for cleaning were introduced (Nordansjö 1988). In the 1950s chemical treatments (herbicides) also came into use to control deciduous trees as this method only cost 25% of the manual methods (Rennerfelt 1948, Häggström 1955, Kardell 2004). As the use of forest products increased, tending the forest became more important (Kardell 2004). Cleaning has been a common practice in Swedish forests since the 1950s (Fig. 1) (cf. Anon. 2004a).
Progress in transport and harvesting techniques continued. The introduction of hydraulic cranes for loading in the beginning of the 1960s, enlargement of the forest road network, and the expansion of waterpower in the rivers changed long distance transportation in favour of trucks (cf. Nordansjö 1988). During the 1960s forwarders were introduced, in 1973 the first complete harvester appeared on the market and gradually the forest workers became machine operators (Andersson 1986, Nordansjö 1988). Motor-manual brush saws were still in use, although the saws, the work practices, and the work organisation were improved (cf. Pettersson 1973). During the 1970s the use of chemical cleaning was under discussion (Ahlén 1971) and it was prohibited in 1984 (Fig. 1) (Anon. 1985).
Figure 1. Total cleaned area in Sweden according to the National Board of Forestry from 1942 to 2002. The area treated during 1942-1961 is presented as 5-years mean values (Vestlund 2001a, Anon. 2004a). * = Chemically treated area was estimated by means of sales during the years 1968-75 (Anon. 1978). The figure for 1976 is the announced area and the presented figures 1977-79 and 1983-85 are the actual area treated with chemicals (Anon. 1979, Anon. 1987). N.B. 1980-82 no area was chemically treated.
0 50 100 150 200 250 300 350 400 450
1940 1945 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000
Cleaned area, total
Cleaned area, total (5-years mean) Chemically treated area*
Attempts to mechanise cleaning were started in Sweden in the 1970s (Berg et al.
1973, Gustavsson & Moberg 1975). A shortage of persons performing cleaning and the ban of herbicides created a cleaning backlog in the beginning of the 1980s (Mellström & Thorsén 1981). Hence, in the late 1980s and beginning of the 1990s some cleanings in industrial forests were performed mechanically (Myhrman 1987, Mattsson & Westerberg 1992). Lindman (1987) found that mechanical cleaning was cheaper than motor-manual cleaning in stands with more than 10 000 stems per hectare before cleaning. However, no more than 20 machines were said to be in use annually during these years (Mattsson & Westerberg 1992, Lidén 1995). The total economy for the machines was poor and it became easier to find persons performing cleaning. These are the principal reasons for the negligible amount of mechanical cleaning today (Ligné 1999). Currently nearly all cleanings in Europe are done with motor-manual brush saws (Ligné 2004).
Motor-manual felling with chainsaws disappeared from large-scale forestry during the 1990s (Lidén 1995, Synwoldt 2001). The harvesting machines and the logistics have been refined, cutting both the time and cost required to deliver round wood to the industrial sites. Unmanned harvesters, in which the forwarder- driver uses a remote control to perform the harvesting are already available, but as the harvester and forwarder cannot operate simultaneously the benefits decreases (Jansson 2001, Thorner 2004). Cleaning, on the other hand, has not progressed in the same way. The average cost for cleaning has increased during the last twenty years compared with logging and regeneration costs (Ligné 2004). Furthermore, the nominal cost per hectare has remained fairly constant since 1990 (Anon. 1992, Anon. 1998, Anon. 1999c, Anon. 2004a), and motor-manual work is laborious.
These may be the reasons why, once again, it is difficult to find cleaners in Sweden (Vestlund 2001b). This has made some cleaning entrepreneurs look for personnel from low-wage countries such as Poland, Russia, and the Baltic States;
and there have also been cases where dishonest persons have received payment without fulfilling their cleaning contracts (ibid.). In Canada too, there are concerns that there will be a lack of cleaners and that costs of cleaning will rise (Anon.
2001a, St-Amour 2004). From the mid-1980s to the late 1990s the intensity of cleanings diminished in Sweden and since 1994 it has become more common to postpone the operation (Nilsson & Gustafsson 1999). A revised Forestry Act came into practice in Sweden in 1994 which deregulated cleaning of stands over 1.3 m of height (Pettersson & Bäcke 1998). There seems to be an unwillingness to invest in the early stage of the rotation-period, even though cleaning has proved to be profitable (cf. Berg et al. 1973, Cain & Shelton 2003, Tong et al. 2005), when cleaning no longer is compulsory. This has led to a renewed interest in mechanised cleaning (e.g. Glöde & Bergkvist 2003, Ligné 2004).
Autonomous unmanned machinery
Initial ideas regarding the automation of forest operations were presented at the Swedish University of Agricultural Sciences (SLU) in 1994-95 (Gellerstedt 1995),
prototypes were built, designed to develop and demonstrate autonomous† control, to study forest environment sensing, and to test mobility concepts (Kourtz 1996).
The project started in 1993 but ended, due to lack of funding, in 1995 (ibid.). In the following years some initial studies regarding autonomous cleaning were made at SLU (Gellerstedt et al. 1999).
The number of other projects that have considered mobile autonomous artificial agents (robots) for forestry operations seems to be limited. Kurabyashi & Asaman (2001) have developed a tracked robot for removing bushes between desired trees in steep terrain, where the desired trees are identified with tags. Canning et al.
(2004) presented a small test vehicle with shaft encoders and ultrasonic sensors that autonomously navigated down a 150m forest path. There is also an ongoing project called: Autonomous Navigation for Forest Machines, which started in 2002 and is a part of a long-term vision to develop an unmanned vehicle that transports timber from the felling area to the roadside, and addresses the problems associated with localisation and obstacle avoidance in forest terrain (Hellström 2002, Hellström et al. 2005). However, research and development regarding autonomous land vehicles for use in other industries can benefit the forestry sector. Interesting studies have for example been made for use in agriculture in recent decades (e.g. Tillett 1991, Marchant et al. 1997, Noguchi et al. 1998, Zhang et al. 1999, Wilson 2000, Have 2002). The agricultural activities addressed have included harvesting, mowing, weed control, and applications of pesticides.
Safety is the most important issue in developing automatic unmanned vehicles, along with reliability, and the unstructured outdoor environment makes it challenging to reach acceptable reliability and safety levels (Tillett 1991).
However, recent development in computer vision and global positioning system (GPS) make such sensors interesting for automatic guidance of agricultural vehicles, especially if data from those two sensor-types could be combined (Keicher & Seufert 2000, Wilson 2000). Durrant-Whyte (2001) states that the necessary sensors, algorithms and methods to develop and demonstrate an operationally viable all-terrain autonomous vehicle already exist. Declining prices for computer power, machine vision, and navigation systems also favour the development of autonomous outdoor vehicles (Keicher & Seufert 2000).
Although, most work in this area is at the pre-commercial stage, except for lawnmowers (Nielsen & Fountas 2002).
Automatic selective cleaning
The cleaning results should not be affected if the operation is automated.
Consequently, in selective cleaning decisions must be made for each stem, either it
† Autonomous refers to something that is independent of others, acting or able to act in accordance with rules and principles of one’s own choosing, from the Greek word autonomos, living under one’s own law (Anon. 1995a). An autonomous robot can adapt to
is a main-stem or it is not, and automating the decision process requires a computer-based program, e.g. a decision support system (DSS).
The mathematical aspects of decision theory usually concern optimisation under constraints. Typical constraints are economical, time or labour. When modelling a decision problem Boman & Ekenberg (1999) states that one has to:
• Compare alternatives with respect to different perspectives, e.g.
environmental, financial, security
• Compare alternatives within each perspective
• Estimate the probabilities that the given status occur, given that a certain act is performed
• Estimate the value of the consequences
An optimal decision requires that the expected utility for each course of action can be calculated (cf. Boman & Ekenberg 1999).
Stem selection in forestry
To perform an optimal cleaning each stem has to be given an explicit grade in comparison to all other stems, which determines if the stem is a main-stem.
However, it is not possible to reliably predict the outcome of a decision to select a certain stem since it cannot be known in advance how this stem will develop in the complex forest ecosystem, i.e. in its interactions with soil, micro organisms, microclimate, plants, animals etc. (cf. Allen & Gould 1986, Mendoza & Sprouse 1989, Gadow & Füldner 1995). Forestry decisions also often concern long time horizons and multiple stakeholders with separate interests, which further complicates the decision-making (cf. Kangas & Kangas 2004). There are, for example, uncertainties regarding how stands will develop after thinning operations (Gadow & Füldner 1995) as well as regarding which tree qualities that will be desired over time. Furthermore, the motives for owning a forest vary (cf. Menger 1934, Hugosson & Ingemarsson 2004) and accordingly the silvicultural goals of different forest owners. Thus, the decision situation in a stand can be characterised by a lack of information and be classed as uncertain, i.e. each action has several possible outcomes for which the probabilities are unknown (Siddall 1972).
Nevertheless, to be able to make individual selections in practice, each stem must be differentiated by some of its characteristics. To be useful to the decision maker, the characteristics should be both comprehensive and measurable (Keeney
& Raiffa 1993). In the decision-making process a cleaner is limited to use the available instructions and information, what he sees, what he has learned, and what he remembers from the previous selections. Kahn (1995) states that decisions made during thinning are partly based on subjective criteria and partly on
Ducey 2001). When one stem is removed, the competition might change and the necessity to remove other nearby stems might disappear (Kahn 1995).
People solve uncertain, ill-structured, problems by the shrewd use of heuristics and at the expense of giving up guaranteed completeness of searches and optimality of the solutions attained (Simon 1995). The solutions applied when solving complex and uncertain decisions are generally good or bad, rather than true or false (cf. Allen & Gould 1986, Gadow & Füldner 1995). To automate the selection of main-stems, computer-based decisions must render acceptable cleaning results. It is not necessary to find the best solution, but to quickly find a sufficiently good one (cf. Daume & Robertson 2000a). Mendoza & Sprouce (1989) replaced the traditional view of optimising the attainment of a given objective with a more practical concept of “satisficing” or attaining a satisfactory level of achievement.
Decision Support Systems
Decision Support Systems (DSSs) are computer-based systems designed to represent and process knowledge in order to support decision-making activities (cf. Holsapple & Whinston 1996). DSS is a broad term and its definition varies, it does not necessarily include artificial intelligence (AI), but DSSs are often used for complex decision-making (cf. Druzdzel & Flynn 2000). AI is that field of computer usage which attempts to construct computational mechanisms for activities that are considered to require intelligence when performed by humans (Partridge 1998). It is also the field of research of human thought processes, to understand what intelligence is. Thus, AI is both a part of computer science and a part of psychology and cognitive science (Simon 1995). AI applications can be either stand-alone software, such as decision support software, or embedded in robots.
Expert systems and Knowledge-based (KB) systems are a kind of DSSs that support the decision-making process in a narrow well-defined area using AI techniques to store and retrieve knowledge and consist of both data and relationships among the data (Mills 1987). Other fields of AI include natural language processing, knowledge representation, machine learning, automatic programming, and pattern recognition (Holsapple & Whinston 1996).
The terms expert systems and KB systems are usually used synonymously (Mills 1987, Holsapple & Whinston 1996). However, expert systems usually perform tasks which normally require a human expert and can involve heuristics, while the term KB systems can also be used when tasks that require detailed knowledge, but not a human expert, are performed (Mills 1987). The two main components of an expert/KB system are the knowledge base, which differs from a database in that it contains executable program code (instructions) and the inference engine, which interprets and evaluates facts, instructions and data in the knowledge base (Waterman 1985).
Thus, automatic selections of stems need some kind of a DSS and that certain characteristics of the stems can be perceived. To develop a cleaning DSS, appropriate objectives for the cleaning must be defined and an appropriate set of attributes and/or rules should be associated to them (cf. Keeney & Raiffa 1993).
However, incorrect decisions can be caused by three kinds of errors (Giarratano &
Riley 1998, Daume & Robertson 2000b):
• Too few attributes are used
• Wrong attributes are used
• Unsuitable interference of attributes
A cleaning DSS should include as many attributes (and rules) as needed to give acceptable result, but as few as possible to make the system simple and fast (cf.
Daume & Robertson 2000a).
The aim of this thesis was to analyse key issues regarding automation of cleaning;
suggesting general solutions and focusing on automatic selection of main-stems.
The objectives of Paper I were to assess forestry requirements (mainly from a Nordic perspective), review available technology, and to suggest a system design for a robot performing selective cleaning in young forest stands. The objective of Paper V was to test if laser scanning could be a possible sensor technique for finding stems in young forest stands.
The objective of Paper II was to assess the explicit and implicit “rules” and ideas used in cleaning in order to facilitate the development of a DSS for selective cleaning. The objectives of Paper III were to develop a DSS for automation of individual stem selections in practical cleaning and to test if it could render acceptable results. Further evaluations of the DSS were made in Paper IV, where the objectives were to compare the cleaning results of experienced cleaners and DSS simulations when “similar” instructions were given, and to assess the usefulness and robustness of the DSS.
Material and methods
A review was made to analyse the requirements a cleaning robot must meet.
Cleaning manuals/instructions, as well as literature regarding mechanised cleaning, forest machines, cleaning-stands, and forest terrain were used to describe typical requests, focusing on Swedish conditions (Paper I and partly used in Paper II). Literature regarding various sensors and localisation techniques, especially for outdoor applications, were used to identify possible solutions to meet the stated requirements. Literature about the detection, identification and classification of trees, path planning, and forest machinery were also used to introduce promising ideas and techniques that could fulfil the stated requirements. (Paper I).
Qualitative interviews were made with thirteen cleaners in 2001 using a semi- structured approach (Patton 1990). The interviewed cleaners, all men, worked in central Sweden and were either entrepreneurs themselves (5) or employed by an entrepreneur (8). Most of their commissions were obtained from industrial forest owners, but some also worked for non-industrial private forest owners (NIPF owners). The cleaners had varying degrees of experience, and the ten cleaners with more than two years cleaning practice were referred to as experienced cleaners. All interviews were taped and then transcribed. The statements the interviewees made mostly concerned themselves, but in a few cases they referred to other cleaners, or cleaners as a group. The interviews concerned:
• Work organisation/situation
• Instructions given to cleaners
• Preferred characteristics of main-stems
• Practical selection
• Foundations of cleaning knowledge/experience
When the interviewees’ answers gave no further information, i.e. when saturation was achieved (Glaser & Strauss 1967), interesting information was clustered and checked for similarities or disagreements. The findings were complementary and no real discrepancies appeared, so the results were presented as a generalised image, a gestalt (Eisner 1998). From this gestalt, further abstractions were made. (Paper II).
The Decision Support System
The abstracted results from the interviews were compared with the literature regarding cleaning instructions, also used in the review, and the conclusions from
these comparisons were used to form a set of basic rules for selective cleaning (Paper II).
A DSS was developed, and the restrictions and attributes included in it were evaluated and improved (cf. Vestlund 2003) in accordance with the variables currently used for representing acceptable cleaning results (presented in Paper II), i.e. the number of stems per hectare, species composition, and percentage of damaged stems. Restrictions for minimum and maximum distances between stems were also included.
The DSS uses three parameters, three thresholds and a “quality criteria”
definition regarding species, diameter, and damage, for selecting main-stems. Four types of damage were used to define damage (see section Field inventory) and the preferred diameter was expressed as area and species specific ranges. The three parameters depend on the purpose of the cleaning. The first parameter is the requested spacing, and concerns the density target and the maximum distance restriction. The squared double-spacing is used to divide the area to be cleaned into smaller parts, here called sections. To reach the density target each section should have four remaining stems, on average, after cleaning. The second parameter is the minimum allowed distance between two stems. This parameter causes the DSS to reject stems if they are situated within this distance from an already selected main-stem. The last parameter is the requested percentage of deciduous stems, and influences the final selection of remaining stems.
In areas where there are too few stems that fulfil the “quality criteria” to reach the density target more stems can be selectable according to two thresholds.
Roughly speaking, the first threshold regards undamaged stems that do not fulfil the “quality criteria” and the second regards damaged stems. The third threshold influences the final selection of remaining stems, and could allow the selection of more, than the average number of four, stems in an area.
The selections rest upon accessible inputs, i.e. data from the current and previous sections, and selections that have already been made are not changed. A decision to save a stem affects forthcoming decisions, especially in surrounding sections, in order to meet the requested over-all targets for the stand. (Paper III).
A field inventory (FI) was conducted in the summers of 2002 and 2003 at two areas near Enköping, two near Skutskär, and two near Jönköping (Fig. 2). The selected areas (Table 1), were in need of cleaning according to cleaning manuals (cf. Anon. 1999a) and the target was to leave approximately 2 500 stems per hectare after cleaning.
Figure 2. Sweden, location of the field inventory areas (Skutskär, Enköping, and Jönköping) and the place where the laser scanning was performed (Bennikebol), see section Laser scanning. The position of the 60th parallel is roughly marked.
The inventoried area was 160 m2 at each location, except the JönköpingPine- area, where it was 224 m2. Retrieved characteristics were: diameter, position, species, and damage. All stems over one cm in diameter at breast height (dbh) were callipered with mm precision. The centre positions of the stems were measured in X and Y-planes at breast height with cm precision. The stems were categorised as Scots pine (Pinus sylvestris L.), Norway spruce (Picea abies (L.) H.
Karst), juniper (Juniperus communis L.), birch (Betula pendula Roth and Betula pubescens Ehrh.; species not separated), or other deciduous. Four damage types, considered automatically measurable, were chosen for defining damage:
1. Double top, where the height of the shorter top was at least 0.5 m and at least 75% of the height of the taller top
2. Leaning stems, i.e. stems having a mean inclination angle larger than 2°
from root to top
3. Stems with crooks, where it was not possible to join the centres of each end Skutskär
Jönköping Bennikebol 60°
4. Stem damage with an area larger than the squared radius at breast height (r2) of the stem.
Stems with damage of types other than those defined were noted as having
Table 1. Stand data from the field inventory (cf. Fig. 2), all stems over 1 cm in diameter at breast height, dbh, were counted and measured
Stand data Enköping Pine1 Enköping Pine2 Jönköping Pine Jönköping Spruce Skutskär Pine Skutskär Spruce
Density (stems per ha) 10000 9875 5893 5500 6188 6938
Proportion of birch stems
(%) 52 33 60 8 2 19
Proportion of “other
deciduous” stems (%) 4 1 0 0 2 36
Proportion of stems with
damage (%) 58 41 65 16 14 60
Proportion of stems with
“undefined damage” (%) 4 1 8 7 5 5
Mean dbh, total (mm) 30 29 40 46 69 36
Mean dbh, coniferous (mm) 42 34 72 47 70 50
Mean dbh, deciduous (mm) 20 19 18 41 30 24
(Papers III and IV).
The DSS was used for a total of seven computer-based cleanings, i.e. simulations.
Firstly, six cleaning simulations with different settings were run (Paper III):
• Changed thresholds
• Increased minimum distance
The “General” simulation was run in accordance with instructions generally given to cleaners (presented in Paper II). The density target was set to 2 500 stems per hectare after cleaning, the minimum allowed distance between stems was 0.5 m, and as a target 10% of the remaining stems should be of deciduous species. To
simulation had the same settings as the “General” simulation but the “Reverse”
simulation started in the opposite corner of the areas.
The following four simulations kept the settings from the “General” simulation except the altered value(s). The “Changed thresholds” simulation increased the importance of reaching the density target by altering the three thresholds. In the
“4000-stems” simulation, the spacing parameter was decreased as the targeted number of stems per hectare was increased to 4 000. The “Increased minimum distance” simulation doubled the minimum allowed distance between stems. In the
“30%-deciduous” simulation the targeted percentage of deciduous stems was increased.
To increase the available data for the six simulations, an old field inventory (OFI) by Gustavsson (1974) with eleven areas was included (cf. Fig. 1 in Paper III). These stands were described as representative Swedish cleaning stands but varied regarding e.g. density, species composition, and height. The utilised areas were 480 m2 at each location. (Paper III). A thorough description of the OFI areas can be found in Gustavsson (1974).
In Paper IV results of the “General” simulation were used again and compared with the cleaners’ results (presented in the section Field “experiments” with cleaners). An “Adjusted” simulation was also made where the preferred dbh range, targets for density, and targets for species mix were altered in accordance with the mean results of the cleaners. The minimum allowed distance between stems and the first two thresholds were set as in the “General” simulation. The last threshold was decreased (as in the “Changed thresholds” simulation). To fulfil the
“quality criteria” coniferous stems were (as before) to be undamaged and within the preferred dbh range. However, both undamaged and damaged deciduous stems within the preferred dbh range were regarded as fulfilling all “quality criteria” in the “Adjusted” simulation. (Paper IV).
Field “experiments” with cleaners
Twelve professional forestry workers were engaged to “clean” the six field inventory areas (cf. Table 1). Forest companies were asked to appoint cleaners known for producing acceptable results, and the participating cleaners, all men, worked in south and central Sweden. They were instructed to select the remaining stems, as in an actual cleaning, considering the desired targets (similar to the General simulation):
• 2 500 stems per hectare
• 10% deciduous stems, “other deciduous” stems should be favoured to increase diversity
• At least 0.5 metres between each remaining stem
• Avoid selecting damaged stems
These experienced cleaners made their choices by indicating on a map the stems they decided to leave in the area. Each cleaner “cleaned” two areas and four cleaners “cleaned” each area. (Paper IV).
Field “experiments” with laymen
Four persons, with little or no forest knowledge, herein called laymen, were given a printed version of the DSS. The laymen functioned as substitutes for actual cleaning robots and were directed to follow the system’s recommendations, with the general settings for the DSS. They “cleaned” the SkutskärPine-area (cf. Table 1). The laymen were given the same damage definitions as the computer, but were allowed to decide for themselves which of the stems were damaged. They indicated on a map the stems they selected, with the aid of the DSS. They also indicated the reasons for their selections on this map. (Paper IV).
Horizontal laser scanning was made in 2001, at two sites in a young forest stand near Bennikebol (Fig. 2). The stand had 7 000 stems per hectare, 50% coniferous and 50% deciduous, and the average height was four metres. Each scanning produced five different layers of data for each pixel: the mirror rotation angle, the laser-plane angle, distance, amplitude, and ambience. The raw data from the scanner was transformed into images, which were analysed to find more or less vertical lines, i.e. trees. Estimates of the height, diameter, and position of these
“trees” were then made. (Paper V).
Means and 95% confidence intervals (CI95) were calculated for the results of the six simulations in Paper III regarding density, proportion of deciduous stems, and proportion of damaged stems. These values and the different simulations’ effects on the diameter were searched for significant differences (p < 0.05). When making pair wise comparisons, Tukey’s test was used to avoid mass significance. The 2- sided F-test was used to test for equal variance. When equal variance could be assumed pooled variances were used, and when two means of small samples with different variances were compared the statistics are referred to as the Behrens- Fisher problem (cf. Everitt 2002). (Paper III).
In Paper IV, treatment effects were analysed with analysis of variance. The experiment was analysed as a randomised block design, with cleaning method as treatment and area as block (cf. Eq. 8 in Paper IV). The three treatments, i.e.
cleaning methods, were manual cleaning and the “General” and “Adjusted”
simulations. In all analyses, there were five degrees of freedom for block, two for treatment, and 10 for error. Treatment means were compared using t-tests with
proportion of deciduous stems, and proportion of damaged stems (cf. Everitt 2002). (Paper IV).
Papers I and V
A cleaning robot needs to find, select, and treat trees in the whole assigned area according to given instructions. The robot must be capable of moving safely within the forest environment, i.e. to navigate and localise itself automatically in a dynamic, non-deterministic and potentially hazardous environment. The vehicle’s size and mass are of importance, e.g. it must be able to work between remaining stems. Furthermore, the robot must be safe for humans and animals, and it should not cause damage to remaining trees. To be cost effective it must be able to work independently night and day more or less throughout the year.
A robot must be able to adapt to various requests from different landowners regarding variables such as desired characteristics of main-stems, number of remaining stems, and percentage of deciduous trees. This makes identification and automatic selection of stems a critical phase in the development of an autonomous cleaning robot. Obstacle avoidance and target identification are identified as the most difficult problems. Machine vision, radar, and laser scanners, and combination of such sensors, are promising techniques for obstacle avoidance, tree identification, and tool control. It is possible to find trees in images produced with data from a laser scanner, and in Paper V the height and diameter of stems with a dbh < 0.05 m was possible to estimate. The above mentioned sensors may also be used to detect humans and animals, but inclinometers are needed to protect the vehicle from damage. Promising navigation and localisation techniques are combinations of GPS and Inertial Navigation Systems. Cost-effectiveness can possibly be reached by using solutions from the multimedia and automotive industries. There are a few relatively small machines currently operated by humans and autonomous systems used in other areas that have potential for further development, and there is also a cleaning tool that could be used. Software components needed for a cleaning robot includes a control system that is responsible for task planning, selection of main-stems, sensor handling, propulsion, and cutting operations and a target handling behaviour that deals with the main task for the vehicle, i.e. moving to a target-tree and cut it. Based on a hybrid of the reactive and hierarchical robot paradigms, an architecture for executing cleaning operations was proposed. The components in the architecture include the following functions: Mission planner, Sequencer, Cartographer, Resource manager, and Performance monitoring. To implement autonomous functions into a robot usable in forestry is however a very delicate task.
The instructions given to cleaners working in an industrially owned forest are often rather stereotyped and NIPF owners usually request a “nice” stand.
Skilful, experienced cleaners state that their selections are made automatically, whereas less experienced cleaners sometimes need to think for a brief moment.
However, the time available to make selections is short, so they are constantly looking around, forming opinions about the stems as they walk through the stand.
A cleaner should on average treat one hectare per day. The information accession range in young forests is restricted by obstructed views to approximately five metres.
Cleaners consider an even stand with many stems of the preferred species, i.e.
many good options, easy to clean. Experienced cleaners think that less experienced cleaners sometimes detect damage too late and that they have a greater tendency than experienced cleaners to leave too many stems per hectare.
Furthermore, inexperienced cleaners need more control and supervision. The contacts the entrepreneurs and cleaners have with the assigners are currently sparse. Previously, cleaners working for forest companies had training-days, in which a cleaner cleaned a designated area and an expert gave him feedback as he worked. Experienced cleaners do not perceive this shortage of information from the owner to be a problem. Nevertheless, cleaners think that they could make better selections if they had more time to perform their work. Time pressures have increased since the payment per hectare is currently about the same as it was ten years ago. The implicit “rules”, i.e. their recurrent statements of preferred and unwanted characteristics, the cleaners use to complete their work differed somewhat from the explicit rules. There were apparent differences in the number of requested stems per hectare and the height allowed for untreated stems and there were also dissimilarities regarding the unwanted characteristics.
The developed DSS selects main-stems by species, position (including distance and density parameters), diameter, and damage. Six simulations were run following the DSS, which showed that the results depend on the initial state of the stands, but generally the requested targets were met in an acceptable way. After the simulations the average density results deviated by -20% to +6% from the target values (Fig. 3, results of FI areas), the amount of deciduous stems shifted towards the target values, and the proportion of stems with defined damaged decreased from 14 - 90% in the initial stands to 4 - 13%. The mean diameter at breast height increased by 40 - 56% in the different simulations and the minimum allowed distance between stems was never violated.
Figure 3. Density results in the FI areas after the simulations (Papers III and IV) and the results obtained by the 12 cleaners (Paper IV). N.B. four cleaners “cleaned” each area and one cleaner “cleaned” two areas, their personal reliability cannot be gauged from this figure.
1500 2000 2500 3000 3500 4000 4500
General Reverse Changed Tresholds 4000-stems Increased minimum distance 30%-deciduous Adjusted Cleaner(s) Cleaner(s) Cleaner(s) Cleaner(s)
(stems per hectare)
EnköpingPine1 EnköpingPine2 JönköpingPine JönköpingSpruce SkutskärPine SkutskärSpruce Target
Changed Tresholds 4000-stems Increased mini- mum distance
The FI areas were used for a comparison of the results obtained by the cleaners and the results of the “General” and “Adjusted” simulations, which revealed that the stand density results were significantly affected by location (Fig. 3). The cleaners’ mean results for the density deviated by -25% to +30% from the instructions, the “General” simulations rendered similar variations (-20% to +40%
in the different areas), and the “Adjusted” simulations deviated by -7% to +13%
from their density targets, which were area-specific. The percentage of deciduous stems varied from 3.6% to 15.0% after the “General” simulation whereas the cleaners’ mean results were significantly higher, varying from 4.0% to 36.7%, and after the “Adjusted” simulation the proportion varied from 4.5% to 37.5%. So, the proportion of deciduous stems was significantly affected by both method and location. The proportion of damaged stems was also significantly affected by both method and location. On average more than 80% of the stems that were selected in one of the simulations were also selected by at least one cleaner, and about a third of the stems selected in the “General” and “Adjusted” simulations were also selected by all four cleaners. The results obtained by the laymen, who were instructed to follow the DSS with general settings, were close to the results of the
The research method must be consistent with the aims of the research. The work of this thesis was explorative in its approach so the methods used were selected and adapted as it progressed.
The motive to perform reviews in the initial stage of this project (Papers I and II) was to assess the knowledge that already exists, although from a new perspective. Reviews are a cost-effective way to acquire knowledge, but can be incomplete when the research originally performed had another scope. For instance, research regarding automatic detection of trees has been made on larger trees (dbh > some 0.2 m) (e.g. Högström 1997, Clark et al. 2000, Forsman 2001, Tarp-Johansen 2001). Therefore, the potential ability of a laser sensor to find trees in young forest stands (dbh < 0.05 m) was investigated in an initial study (Paper V). The material in Paper V was limited, since the financial resources were restricted, but the study indicates that laser sensors may be used to find and measure small trees.
The review showed that cleaning manuals/instructions could not be used as a single source for automating the selection of main-stems as they are both too general and too detailed and mostly concern the end results. Still, cleanings are performed, so there seems to be tacit knowledge, i.e. persons involved in cleaning seem to have an implicit understanding of how to acquire relevant information, make selections and proceed within the stand. So, to gain an understanding of how practical cleanings are performed qualitative interviews were made. When humans are dealing with complex situations, they use only a few data to make decisions.
Magnusson (1978) found that a typical doctor used from one to five out of ten available, considered important, values for making decisions about how much blood a patient should be given during a transfusion. Another test showed that a psychologists ability to predict a students result on a test rose when the available variables went from two to four, but with five or six variables the prediction ability was stabilised or even dropped (ibid.). The knowledge of a human expert is often heuristic in nature, based on useful ”rules of thumb” rather than absolute certainties (Cawsey 1997), see also the section Automatic selective cleaning.
Heuristics can provide valuable shortcuts that can reduce both time and cost thus there is reason to explore the human heuristic search techniques as a source of ideas for developing “intelligent” computer-based systems (cf. Simon 1995, Giarratano & Riley 1998).
Semi-structured interviews were used as they can be modified over time, to focus attention on areas of particular importance or to adapt to changing circumstances or new understandings (Bliss & Martin 1989, Patton 1990).
determine if qualitative results are valid (Eisner 1998). To overcome this limitation, cleaners with varying degrees of experience of forest work were interviewed (cf. Patton 1990).
To be able to automate the selection process a computer-based DSS was developed, which performs the same functions as a cleaner, but does not mimic humans. The reason for not including heuristic rules in the DSS was that the cleaners’ implicit rules were unobtainable and perhaps wrong (discussed in section Prospect of automatic selection of stems). The advantage with a straightforward design, where the complexity of the task is reduced, is that once the attributes have been captured the decision process is deterministic and uncomplicated. A potential disadvantage is that the selections might be inferior.
However, the attributes used in the DSS were selected since cleaners and cleaning manuals/instructions mention them, as discussed in Paper II, and because it should be possible to detect them automatically, according to the results presented in Papers I and V. Although, attributes like species and damage require thorough and careful description in order to be determined automatically. The dbh ranges were used to increase the mean diameter and to reject stems with the largest diameter, i.e. to decrease the disparity in dbh and thus create a more uniform stand, which is a usual request according to Paper II. Another aspect of the attributes was that they had to be possible, at this point, for a human to measure. In the future, when information regarding the stems is provided through sensors, attributes like “amplitude” or IR-light etc. could be used. To have a practicable system, retrieved attributes and decision-making was limited to small-scale areas, as distant information usually is unobtainable when cleaning due to the restricted view (according to Paper II). The DSS was designed to have the ability to allow the selection of more stems than the average four, to compensate for sections not yet visited which could have a lack of desirable stems. However, when the settings in the DSS are adjusted to the initial stand and in accordance with the landowners requests this ability should be reduced since it can render overall results deviating from the target.
The field inventory was conducted to enable adjustments to be made to the DSS and thereafter to evaluate it (Papers III and IV). The DSS was developed to suit conventional Swedish cleanings stands, i.e. cleanings of stands at some three or four meters of height with a predominance of coniferous stems remaining after cleaning (cf. Brunberg 1990, Varmola & Salminen 2004, Ligné et al. 200X). A usual request at company owned forest is to have 2 500 stems per hectare after cleaning of which 10% should be deciduous (Paper II), and the FI areas were selected accordingly. It should be noted that dead trees were not measured as it was supposed that the DSS would be able to sense whether trees were living or dead. One of the reasons for making this field study was to acquire information about the stems’ position with cm-precision. The stems’ positions were needed to evaluate the DSS, i.e. to perform simulations. To enlarge the material for the simulations an old field inventory from the 1970s was included. This study had dm-precision regarding the stems’ positions and another approach was used to
available data and made it rather diverse (cf. Bergstrand et al. 1986). The stands are examples of Swedish forests, but it was not possible to determine their representativeness regarding the proportion of damaged trees and diameters, since the damage definitions were study-specific and no other larger studies considering the diameter were found. The density and proportion of deciduous stems seem to be acceptable according to Table 2. Furthermore, the FI areas were appointed by silvicultural managers as being typical cleaning stands of their organisations. The OFI areas were described as being representative cleaning stands in Sweden (cf.
Gustavsson 1974), which seems correct according to Table 2, although it cannot be confirmed.
Table 2. Mean values for the density and proportion of deciduous stems in stands before cleaning as described in different studies
FI and OFI areas (17 areas)
Inventory of 1997 (457 stands)*
Vestlund 2001 (5 areas)
Swedish national inventories Mean density
(stems per hectare)
(dbh > 0.01 m) 8 000 (stems over 2
cm at cut- height)
12 640 (all stems in 2
areas and stems above breast height in
9-12 000† (all stems,
inventory made in 1993-96) Proportion of
46% just above 50% 55% 62%‡
Mean height (m)
(OFI areas) 2.52 4□ (newly
* Pettersson & Bäcke 1998
† Values from the Swedish National Forest Inventory presented by Pettersson & Bäcke (1998)
‡ Anon. 2002
□ Nilsson & Gustafsson 1999
There are a large number of desirable and undesirable attributes of a tree (see Paper II) but the way individual selections of for example two “comparable” stems affect the stand development is not possible to exactly predict. Furthermore, the assigners’ actual requests are in some cases unsaid (see Paper II). This limits the possibility to make optimal decisions or an exact validation of the DSS. Thus, three other methods were used to evaluate the system. First, six simulations with different settings were run in the 17 areas, to evaluate how these settings influenced the results, and to assess the acceptability of the results (Paper III). To further evaluate the DSS a comparison was made with results of twelve cleaners (Paper IV). Kahle (1995), Zucchini & Gadow (1995), and Daume & Robertson (2000b) have used this method of evaluation in similar studies. A further intention of the comparison was to demonstrate differences between the cleaners, as this would imply that even if the DSS does not meet all of the targets it could be at least as good as a cleaner. Using a non-destructive method enables comparison of
verified. The results of the cleaners were also used to alter the settings and make an “Adjusted” simulation, in order to assess the possibility to adapt the DSS to their selections. Thirdly, laymen used the DSS for a cleaning to illustrate the systems potential utility for a robot, they functioned as substitutes for actual robots, and as a training-tool (Paper IV). However, it should be noted that this last method reflects not only the DSS’s potential as a training-tool, but also my ability to rephrase the computer-rules into a “language” that people can understand. Still, the use of three different evaluation methods indicates the DSS’s capability to reach acceptable results in a variety of situations. To further evaluate the DSS, and to refine the system if the assigners do not accept the system’s selections, more simulations could be made according to alternative instructions given by the assigners. There is also an option of making comparisons with selections made by silivicultural researchers. The selections and results of the DSS could also be evaluated with a growth simulator, as the one presented by Fahlvik et al. (200X).
However, that kind of evaluation is currently of less relevance, as the aim of this DSS was to be operational and to produce results accepted by assigners and comparable to cleaners.
Prospect of a cleaning robot
The importance of forestry is considerable for Sweden as a whole, but its importance for individual Swedes has changed. The first paragraph of Sweden’s Forestry Act states that: “The forest is a National resource. It shall be managed in such a way as to provide a valuable yield and at the same time preserve biodiversity. Forest management shall also take into account other public interests” (Anon. 1995b). The perceived value of the forests has shifted from purely economical to become a recreational value for most people (cf. Hörnsten 2000). The multiple purposes of forests might be one reason that the production of timber is not usually regarded by ordinary people as the business activity it really is. However, continuous progress is needed in forestry to maintain its profitability and to meet the demands of the surrounding world.
Thus, is it likely that forestry, which has progressed from manual through motor-manual to mechanical operations, will be automated in the future? Will cleaning robots become practical realities? Perhaps, but as discussed in Paper I, there is still a long way to go.
1. An automated system must seem to render such advantages that it is interesting to develop.
2. The robot must be able to operate in forest environment, in real-time.
3. It must be possible to integrate the robot in a silvicultural system at large, i.e. it must function and operate in organisations of people and other machinery.
4. The whole work operation for cleaning must be described from a robot’s perspective.
Since systems for cleaning (manual, motor-manual, mechanical and chemical) already are available, it is necessary for a new system to appear as more attractive by e.g. decreasing the workload, improving the quality, and/or decreasing the costs (Nåbo 1992). Changes in regulations may also force a system change, as can be seen in Canada, where the use of herbicides is already prohibited in Quebec and being debated in other areas (cf. Ryans & St-Amour 1996, Pitt et al. 2000, Anon.
The most challenging problems in developing a robot usable for cleanings are probably obstacle avoidance and target identification in the forest environment.
Combinations of machine vision, radar, and laser scanners appear to be promising solutions. Necessary sensors, algorithms, and methods to develop and demonstrate operationally viable outdoor autonomous vehicles already seem to exist (cf.
Durrant-Whyte 2001). However, it is not possible to state at this early stage of the development of robots whether they will be able to deliver reductions in costs or better work environment for the personnel. A cleaning robot would be quite expensive considering all its technical equipment. The economy of it is based on the prospective use of an inexpensive base-machine and off-the-shelf equipment/solutions from other industries/areas of research. Design issues regarding e.g. the robots propulsion, size, and tools should be easier to solve, as there already exists fully operational forest machinery. For example, in 1995, the forest machine company Timberjack presented a walking harvester (cf. Anon.
In order to reduce labour costs, i.e. costs of operators, the robot should solve common problems automatically, although it must be able to call for assistance when it encounters unusual difficulties. This indicates that its productivity per hour could be lower than that of conventional methods. In a normal week a cleaner works 40 h. If the robot could work unattended throughout the week, both night and day, it would have 168 hours available, i.e. it could be four times slower and still compete with a cleaner if its cost per hectare was equal. In comparison with traditional forest machinery, which is used up to 3000 h per year, a robot working continuously has 8760 h available per year. However, transportation between sites, situations with faults or alarms, and maintenance would decrease the productive time and the cost per hour might be higher for the robot.
A robot would turn the personnel into operators, releasing them from heavy workloads, but might increase their mental burden (see also the section Changing the system). However, the current cleaning situation is not sustainable in the long- term as it is not possible to expect that there will always be persons available to work for a low remuneration. Thus, there is or at least will be a need for change.
Both time and effort is needed in order to develop a cleaning robot, and forest machine manufacturers differ in their opinion regarding whether or not robots will become reality in the future (cf. Jansson 2001, Thorner 2004). To promote automation in the forestry sector an expressed desire from the forest companies is
operations. There is yet a long way to go before fully operational robots for forestry are available. To implement autonomous functions into a robot used in complex environments, like the forest, is difficult since such a robot not only have to build maps and find its way through the unknown forest but also perform assignments, e.g. cleaning, when it has reached its destination (cf. Uhlin &
Johansson 1996). The answer to this general problem in robotics is hybrid systems that combine hierarchical and reactive components (cf. Blackmore et al. 2002) and in Paper I a proposed system design for a cleaning robot with hybrid and reactive behaviours is presented. Murphy (2000) states that in the hybrid deliberative/reac- tive paradigm the robot first plans (deliberates) how to best compose a task into subtasks (also called mission planning), and then what are the suitable behaviours to accomplish each subtask, etc. Then the behaviours start executing as per the reactive paradigm (ibid.). In the reactive paradigm all actions are accomplished through behaviours. Behaviours are a direct mapping from sensory inputs to motor outputs that are used to achieve a task (cf. Brooks 1986, Murphy 2000). Thus, it seems more likely that an autonomous shuttle system, operating in pre-planned paths, replacing the forwarder will appear first (Thorner 2004).
Changing the systemCurrent system
The current cleaning system often involves more than one person, especially when done professionally. NIPF owners usually do the cleaning themselves (Lidestav &
Nordfjell 2002), but otherwise cleaning entrepreneurs or cleaners do it. Their equipment consists of: cleaning-saw, blades, files for the blades, fuel, and protective clothing (Fig. 4). Public and forest roads are used for transportation, and cars or vans are needed to transport the personnel and their equipment. Such transports should be quite easy to organise, but the transportation of rest-huts needs some planning.
A cleaning operation typically consists of the following tasks:
• Planning in the office
• Planning in the forest (e.g. marking borders with paper-strips)
• Actual cleaning
• Maintenance of the saw
• Follow-up work
The cleaning system relies on back-up systems including mechanical service and the provision of spare parts, road maintenance and construction, as well as planning and administration.
The way cleaning operations are organised might not be changed, but the assignments will change with an automated system. The cleaners will become operators and their education will need to be modified to handle the machine. The operators will be in charge of the (probably quite expensive) robot(s) and must be capable of controlling the robot as well as perform daily-maintenance of it. The operators will also plan and follow-up the work. If one person does not possess all the required knowledge, a team of persons with different skills could work together, and perhaps learn from each other. Team-work is frequently used in changing environments (Scott 2003). However, team-building is a delicate task, and needs thorough analysis. Even when the 2-man chain saws were introduced in Sweden Zimmermann (1948) stated the importance of selecting operators with great care. Similarly, when mechanical cleaning was introduced at the forest company Stora, the importance of selecting team-members was stressed (Tosterud
& Bergqvist 1990).
Current cleanings involve little pre-planning (Vestlund 2001b). If the work was done with a robot more planning might be needed, and it would certainly have to be computerised. For example, the area selected for treatment could be marked out from the surroundings and the route for the machine to take could be planned in advance on digitised maps. There might also be areas that are not suitable for a cleaning robot, which would have to be treated conventionally with cleaning saws.
Map information used by forest companies is already usually digitised and could probably be transformed for use in an autonomous process. However, cleaning with robots might require more rigorous planning, e.g. to make the correct settings in order to have acceptable and requested results. The robot would produce results according to the given instruction (see also the section Automatic selections with the DSS). If the robot should be unable to follow the given instructions it should stop and call for assistance.
Robots would probably be transported on small trucks and their transport might